What it is
Adapts video MLLMs from autoregressive generation to direct candidate scoring for temporal grounding and multiple-choice video QA, with reusable video-prefix inference. Instead of generating timestamps or answers token by token, it scores temporal intervals and answer options with one shared head over a pretrained video MLLM, then reuses the same encoded video via cached or joint inference for many queries.
Key results
- 67.7 avg grounding mIoU, +10.8 vs Qwen3.6-35B-A3B backbone
- 65.3 avg grounding mIoU for 4B, +6.7 vs Qwen3.5-4B backbone
- Over 100x QA inference speedup vs Thinking baselines at comparable scores
Direct structured decisions can preserve MLLM semantic flexibility while making repeated video judgments far faster and often more accurate.